The Average That Lied to Everyone
Every month, the leadership team at a fast-growing food delivery company gathered around the same slide. It showed a single number in bold, reassuring font:
Investors were happy. Marketing loved it. Engineers nodded and moved on. Nothing looked wrong — until customers started leaving.
Complaints told a very different story: some orders arrived in 15 minutes, others in 90. Support teams were overwhelmed. Social media sentiment dropped. Yet the number on the dashboard stayed calm, stable, and convincing.
This is the story of how the average lied — not because math is evil, but because humans kept asking it questions it was never meant to answer.
The Seduction of the Mean
The arithmetic mean is seductive because it feels objective. You add everything up, divide by the count, and receive a single, clean answer. It feels like truth distilled into one number.
But the mean does not describe reality — it summarizes it. And summaries lie when distributions are uneven.
In the delivery company’s case, most orders clustered between 20 and 35 minutes. But a small fraction — delayed by rain, traffic spikes, or staffing shortages — took over an hour.
Those long delays pulled the mean upward just enough to hide the real experience of most customers, a phenomenon deeply connected to skewed distributions, as explained in distribution visualization examples.
Mean vs Median: Two Very Different Stories
When an analyst quietly calculated the median delivery time, the number was shocking:
That meant half of all customers received their food in under 24 minutes — far better than the advertised average. So why were people angry?
Because the median hides pain at the extremes. It ignores how bad the worst cases are.
The mean hid the tail. The median ignored it. Both were incomplete truths.
This distinction mirrors the same logic used in income inequality analysis, where mean income skyrockets while median income stagnates — a statistical illusion explored in aggregate vs individual analysis.
The Shape That Changes Everything: Skewed Distributions
If delivery times followed a neat bell curve, the mean would be reasonable. But real systems are rarely symmetric.
They are skewed — often heavily.
Right skew meant rare but extreme delays dragged the average upward. Left skew would have hidden exceptional performance.
Skewness is not a math detail; it is a signal. Ignoring it is like ignoring turbulence because the plane’s average altitude is stable.
This exact failure appears repeatedly in data analysis, including clustering mistakes discussed in cluster overlap analysis.
Outliers: Noise or Warnings?
Leadership initially labeled the long delays as “outliers.” Someone even suggested removing them from reports.
That would have made the average look fantastic. It also would have destroyed trust.
Outliers are not always noise. Often, they are the only data points telling the truth about system limits.
This is why blindly removing extremes — a common preprocessing mistake — can lead to catastrophic decisions, as shown in outlier removal case studies.
When Averages Shape Policy
The company redesigned staffing based on average demand. But demand spikes did not follow averages. They followed events: rain, holidays, paydays.
Average-based planning under-allocated resources exactly when they were needed most.
This mirrors failures in public systems — from hospital capacity planning to traffic management — where mean-based assumptions collapse under real-world variance.
The same logic is discussed in operational risk analysis in risk modeling frameworks.
The Machine Learning Parallel
The data science team trained a prediction model to minimize mean squared error. It performed excellently — on average.
But worst-case predictions were terrible. Exactly the customers most likely to churn were the ones the model failed.
This is the optimization illusion: minimizing average loss while maximizing real-world damage.
Loss functions behave just like averages — they flatten experience into a single scalar, as explained in loss function trade-offs.
Why Humans Love Averages (and Why That’s Dangerous)
Averages reduce cognitive load. One number is easier to communicate than a distribution.
But simplicity trades accuracy for comfort.
Executives want clarity. Dashboards want symmetry. Slides want bold fonts. Reality refuses all three.
This psychological bias mirrors how people misinterpret probabilities, a theme recurring in decision interpretation errors.
The Moment Everything Broke
A viral post showed a customer waiting 94 minutes for cold food. The average didn’t matter anymore.
The company finally plotted the full distribution. For the first time, everyone saw the long tail.
Silence followed.
What Fixed the System
They stopped asking the average to tell the whole story.
Dashboards now showed:
- Median - 90th percentile - Worst-case scenarios
Staffing was aligned to peaks, not means. Models were evaluated on tail performance.
This shift mirrors best practices in robust modeling discussed in data evaluation strategies.
The Real Lesson
The average did not lie maliciously.
We lied to ourselves by asking it to explain a world shaped by extremes.
Averages are summaries, not guarantees. They describe no one’s actual experience — especially in systems where failure matters more than success.
Final Thought
If one number feels too comforting, it’s probably hiding something.
No comments:
Post a Comment